Informatica

In this paper an exploratory classification, so called open set problem, is investigated. Open set recognition assumes there is incomplete knowledge of the world at training time, and unknown classes can be submitted to an algorithm during testing. For this problem we elaborated a theoretical model, Double Probability Model (DPM), based on likelihoods of a classifier. We developed it with double smoothing solution in order to solve technical difficulties avoiding zero values in the predictions. We applied the GMM based Fisher vector for the mathematical representation of the images and the C-SVC with RBF kernel for the classification. The last contributions of the paper are new goodness indicators for classification in open set problem, the new type of accuracies. The experimental results present that our Double Probability Model helps with classification, the accuracy increases by using our proposed model. We compared our method to a state-of-the-art open set recognition solution and the results showed that DPM outperforms existing techniques.